Skip to main content
Setup is an authoring workflow outside the visitor request path. Its job is to turn website evidence and team context into a reviewable draft, not to silently redesign production during a visit.

​
Inputs and stages

The setup pipeline reads public pages and their structure/styles, extracts company/brand context, proposes audiences and opportunities, generates approved-target alternatives, validates configuration and saves draft work. Goal discovery and design/placement review support the launch process. The setup directory separates public fetching, site reading, selector stability, preview context, navigation, style extraction, generation prompts, assembly, validation and checkpoints. Use those boundaries when extending it; avoid putting browser delivery logic into a research job. External page text and documents are data. Generated claims must remain grounded in sources; generated persona hypotheses are not verified facts about real visitors. Review typography, responsive behavior, links, preserved native controls and factual claims on the actual site.

​
Jobs and recovery

Active job execution and cancellation controllers live in the process. Saved run state/checkpoints live with durable records. Optimistic checks protect edits made while work is underway. The current startup path calls resumeInterruptedSetups(); recovery depends on the saved phase/checkpoint and must be inspected, not assumed complete after a restart. The service can start in standby for a controlled handoff. Background activation is a separate operation from accepting HTTP traffic. Do not run multiple active setup/catalog owners against the same workspaces. Regeneration can replace draft work and incur authoring-model usage. Retry/cancel/regenerate are different operations; use the returned run state and preserve human edits. A preview-only demo stays unpublished even if setup reports ready.

​
Variant editing

The variant pipeline constructs context from current company/brand guidance, native references and memory, generates/revises a bounded component, sanitizes imported HTML, and validates the saved result. Preview context helps render against the site’s existing design language; it is not arbitrary runtime CSS permission. The server importer is stricter than the browser allowlist. A successful generation can still produce warnings or a poor layout. Read the saved result and preview it after sanitization.

​
Authoring memory

Website memory uses a scoped file-like structure backed by the configured memory integration. Generated files mirror research/brand/personas and are read-only to ordinary note edits. Agents can add durable notes, evidence-backed learnings and dated results snapshots. Search first, then read the actual source before quoting a result. Memory helps the next authoring decision. It is not sent to Jev as live retrieval, does not prove a hypothesis, and does not automatically publish a better-performing design. “Validated” learnings require appropriate live holdout evidence.

​
Preview surfaces

The Chrome extension can run scoped drafts on the real website. Snapshot previews provide a script-controlled rendering context and have different interaction fidelity from the live page. Manual overrides show a selected design; automatic previews exercise selection with a supplied journey. Use both, and identify which mode produced evidence. Source: setup modules, variant modules, memory modules, extension.